YARN-3251. Fixed a deadlock in CapacityScheduler when computing absoluteMaxAvailableCapacity in LeafQueue (Craig Welch via wangda)

hadoop-yarn-project/CHANGES.txt

@@ -619,6 +619,9 @@ Release 2.6.0 - 2014-11-18
     identifiers to be tampered and thus causing app submission failures in
     secure mode. (Jian He via vinodkv)
 
+    YARN-3251. Fixed a deadlock in CapacityScheduler when computing 
+    absoluteMaxAvailableCapacity in LeafQueue (Craig Welch via wangda)
+ 
   BREAKDOWN OF YARN-1051 SUBTASKS AND RELATED JIRAS
 
     YARN-1707. Introduce APIs to add/remove/resize queues in the

hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/main/java/org/apache/hadoop/yarn/server/resourcemanager/scheduler/capacity/LeafQueue.java

@@ -115,6 +115,8 @@ public class LeafQueue extends AbstractCSQueue {
   
   private final QueueHeadroomInfo queueHeadroomInfo = new QueueHeadroomInfo();
   
+  private volatile float absoluteMaxAvailCapacity;
+  
   public LeafQueue(CapacitySchedulerContext cs, 
       String queueName, CSQueue parent, CSQueue old) throws IOException {
     super(cs, queueName, parent, old);
@@ -134,6 +136,10 @@ public class LeafQueue extends AbstractCSQueue {
     float absoluteMaxCapacity = 
         CSQueueUtils.computeAbsoluteMaximumCapacity(maximumCapacity, parent);
         
+    // Initially set to absoluteMax, will be updated to more accurate
+    // max avail value during assignContainers
+    absoluteMaxAvailCapacity = absoluteMaxCapacity;
+
     int userLimit = cs.getConfiguration().getUserLimit(getQueuePath());
     float userLimitFactor = 
       cs.getConfiguration().getUserLimitFactor(getQueuePath());
@@ -720,8 +726,18 @@ public class LeafQueue extends AbstractCSQueue {
   }
   
   @Override
-  public synchronized CSAssignment assignContainers(Resource clusterResource,
+  public CSAssignment assignContainers(Resource clusterResource,
       FiCaSchedulerNode node, boolean needToUnreserve) {
+    //We should not hold a lock on a queue and its parent concurrently - it
+    //can lead to deadlocks when calls which walk down the tree occur
+    //concurrently (getQueueInfo...)
+    absoluteMaxAvailCapacity = CSQueueUtils.getAbsoluteMaxAvailCapacity(
+      resourceCalculator, clusterResource, this);
+    return assignContainersInternal(clusterResource, node, needToUnreserve);
+  }
+  
+  private synchronized CSAssignment assignContainersInternal(
+    Resource clusterResource, FiCaSchedulerNode node, boolean needToUnreserve) {
 
     if(LOG.isDebugEnabled()) {
       LOG.debug("assignContainers: node=" + node.getNodeName()
@@ -1012,12 +1028,6 @@ public class LeafQueue extends AbstractCSQueue {
         computeUserLimit(application, clusterResource, required,
             queueUser, requestedLabels);
 
-    //Max avail capacity needs to take into account usage by ancestor-siblings
-    //which are greater than their base capacity, so we are interested in "max avail"
-    //capacity
-    float absoluteMaxAvailCapacity = CSQueueUtils.getAbsoluteMaxAvailCapacity(
-      resourceCalculator, clusterResource, this);
-
     Resource queueMaxCap =                        // Queue Max-Capacity
         Resources.multiplyAndNormalizeDown(
             resourceCalculator,